GRACE water-storage deficits and surpluses decayed at the same rate on 2,243 land cells in 2002-2026; both stayed longer after extremes than linear precipitation models predict

Question: do terrestrial water-storage deficits decay more slowly than surpluses after the precipitation response is removed? Approach: we averaged CSR GRACE/GRACE-FO RL06.3 mascons (2002-04 to 2026-06) and GPCP v2.3 precipitation to 2,243 ice-free land cells of 2.5 degrees. We removed a monthly climatology and a linear trend. On each cell we fit a precipitation-forced storage model with separate decay coefficients for deficits (a-) and surpluses (a+). We compared the results with 1,000 simulations of a symmetric linear model that use the observed precipitation and block-resampled residuals. Results: the global median a- minus a+ was 0.005 (null 95% range -0.025 to 0.012, p = 0.22). A noise-robust instrumental estimate, a kink at the cell median, a state-dependent precipitation gain and half-recovery times also gave no difference from the null. The test detected an injected asymmetry of 0.05 in all 100 simulations. On 2015-2026 data, the asymmetric model did not forecast better than the symmetric model (median error change 0.0%). After peaks larger than 1 standard deviation, both deficits and surpluses kept 0.08-0.09 of the peak more than a one-reservoir model predicts at 5-12 months (p = 0.001 each), with no sign difference. A two-time-scale model reduced this excess to 0.04-0.05 and decreased out-of-sample error in 77% of cells. Conclusion: at these scales, storage recovery is sign-symmetric but has more than one time scale. Limitations: 2.5 degree monthly scale, precipitation as the only forcing, short records, a trend-adjusted baseline, and simple glacier masks.

GRACE · terrestrial water storage · drought recovery · persistence asymmetry · GPCP · linear reservoir · hydrological memory · null model · threshold autoregression